[FBumann]

Prompt: APply the changes, fixes and refactorings you find reasonable as a stacked PR as you finish the review

Note

The following content was generated by AI.

Stacked on #2022, from a review of it. A block that drops a row with no term now renumbers its quadratic share too, and the omitted-row count adds up. Both bugs are on main as well.

What changed, what was verified, what was not done

Commits

1. test — two strict xfails that fail on the perf(engine): a model with many variables and constraints builds about 15% faster #2022 head (611d936):
   • test_a_quadratic_row_keeps_its_own_pair_when_a_row_before_it_is_not_built: k * p * p <= 10 over g=[a,b,c] with k only at b, c. Row a has no term and is dropped. The quadratic share kept rows 1, 2 and was then filtered to the renumbered rows 0, 1, so qmatrix was [(1, 1, 1, 1.0)]: row c held b's pair and b held none.
   • test_rows_a_restriction_removed_and_rows_with_no_terms_are_counted_together: one row removed by a restriction and one with no term reported rows_not_built = 1, not 2. _drop_termless_rows used = where _labelled uses .get(name, 0) +.
2. fix — _drop_termless_rows renumbers the quadratic share with the same map, and adds to the omitted count. The filter in _block goes, because every row the quadratic share holds is in kept. The markers come off, and what was wrong moves into the docstrings.
3. chore(bench) — bench/profile_build.py also wraps pl.collect_all, and patches labels.frames rather than labels.frame, which the build no longer calls. Before this, the batched passes of perf(engine): a model with many variables and constraints builds about 15% faster #2022 were not timed at all. On network/xs they now show as collect_all of N: … entries under _build_constraints, _build_variables and frames.
4. chore(engine) — a constraint's labelled block stays an in-memory frame through the build and moves onto the model's rows with _moved, as a variable's does. Before, it was a lazy plan (filter, replace_strict, + offset) that ran again on every dual or activity read.
5. The changelog line, and a merge of perf(engine): a model with many variables and constraints builds about 15% faster #2022's new head 44423fb, which merged main and reworded perf(engine): a model with many variables and constraints builds about 15% faster #2022's changelog line. The merge touches no file under src/ or tests/. After it, the touched test modules were re-run.

Mutation table, by hand (the tree had to be committed first; restored with git checkout --, __pycache__ dropped before and after):

mutation	result
constraint offset _moved(frame, 'row', offset).lazy() → frame.lazy()	caught, 5 failed beyond the 2 xpress failures
quadratic share not renumbered	caught by the new quadratic test (it was the xfail)
+= back to = on the omitted count	caught by the new diagnostics test (it was the xfail)

Gates, in a uv venv on polars 2.0.0 (no pixi in this container): ruff check . and ruff format --check . at 0.16.9, pyrefly check on src/ at 1.3.1 (0 errors), pytest -n auto tests: 5446 passed, 657 skipped, 4 xfailed, and 2 failed. The 2 failures are the [xpress] cases, which fail the same way on 611d936 because the xpress package is not installed here. bench/test_harness.py: 396 passed. 1 failed (test_an_ab_refuses_to_measure_a_checkout_it_did_not_import) because specsolve is on PYTHONPATH here and not installed. docs-test was not run.

Not done, from the same review:

• Peak memory: _build_constraints holds every constraint's raw term table at once, and run now stacks them all before one collapse. Not measured here; it needs a memory ladder at network/l before a change.
• The refuse_uncovered guard still has no probe. The door refuses a null parameter value, and a missing coordinate is caught first by refuse_short_constants, so I found no input that reaches it. It may be dead code, which is a decision for the author.
• The first error now follows pass order, not file order. No change.
• declared_height, refuse_null_constants and refuse_short_constants still run once per constraint. Batching them needs its own measurement.
• profile_build attributes attach to setup: the module attribute it patches is not the one the caller imported. This is also the case on main.

Departures from Part 2: none.

🤖 Generated with Claude Code

https://claude.ai/code/session_013DSGmeGQfUxE4gyNQy94qG

[read-the-docs-community]

Documentation build overview

📚 specsolve | 🛠️ Build #35116420 | 📁 Comparing 54f4059 against latest (af1dac7)

  🔍 Preview build  

2 files changed

± about/benchmarks/index.html
± about/changelog/index.html

[codspeed]

Merging this PR will improve performance by 16.23%

⚡ 9 improved benchmarks
✅ 95 untouched benchmarks
⏩ 307 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_emit[network-s-specsolve-lp]	34.5 MB	28.1 MB	+22.59%
⚡	test_emit[network-s-specsolve-in-memory-lp]	34.5 MB	28.1 MB	+22.58%
⚡	test_window[network-s-specsolve-highs-one]	21.9 MB	18.6 MB	+18.23%
⚡	test_window[network-s-specsolve-highs-values]	21.9 MB	18.6 MB	+18.23%
⚡	test_emit[network-s-specsolve-highs]	21.9 MB	19.4 MB	+13.34%
⚡	test_emit[network-s-specsolve-in-memory-highs]	21.9 MB	19.4 MB	+13.34%
⚡	test_window[network-s-specsolve-highs-coefficient]	21.9 MB	19.5 MB	+12.78%
⚡	test_window[network-s-specsolve-highs-cold]	21.9 MB	19.5 MB	+12.78%
⚡	test_window[network-s-specsolve-highs-shape]	21.9 MB	19.4 MB	+12.78%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing claude/gallant-wright-89731k (54f4059) with claude/batched-build (44423fb)}

Footnotes

1. 307 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[FBumann]

Note

The following content was generated by AI.

Folded into #2022 as a fast-forward of claude/batched-build onto this branch's head 54f4059, to keep the change in one PR. The fix's changelog line now links #2022. On #2022's branch with pixi, xpress and Gurobi installed, pixi run check passes (5599 passed, the two [xpress] cases included), and so does pixi run test-bench (411 passed).

---

Generated by Claude Code